A data-driven approach to the processing of sniffer-based gas emissions data from dairy cattle

• A pipeline for the stepwise processing of raw data from methane emission sniffers, including instrument monitoring, time synchronization, linking with animal IDs, background concentration estimation, and operation with data on daily and weekly methane records from dairy cows was developed and validated. • Objective criteria for the selection of valid data were established. • The obtained concentration phenotypes were found to be moderate to highly repeatable. “Sniffers” record methane (CH 4 ) emissions from cows visiting milking robots, providing gas concentration data. These instruments have infrared carbon dioxide (CO 2 ) and CH 4 sensors, an air pump, and a data logger. In this study, a process for the synchronization of sniffer emissions data with cow identification (ID) data and records from automatic milking systems (AMSs) was developed. The process enables the extraction of gas phenotypes for genetic analysis. It involves the calculation of intermediate control variables to account for time drift in data loggers, sensor calibration drift, and background concentration fluctuations, and the condensation of data from each milking visit into a single datapoint. The process was developed and assessed with research station data from three groups of approximately 70 cows, each accessing one AMS unit over a 2-month period. Raw emissions data, including clock times, from CH 4 and CO 2 channels were recorded every second. They were synchronized with the AMS data using specific events occurring in the CH 4 or CO 2 channel at the beginning or end of each milking event. The synchronized data were divided into non-milking (baseline, ambient gas concentrations) and cow ID–linked milking (cow emissions) sets. The non-milking periods varied in duration from a few seconds to hours, and some were interrupted by unrecorded events. Baseline values were extracted after the filtering of non-milking period data against unrecorded events (e.g., washing, feed-only sessions) and the use of a small fractile as the baseline estimate. At the beginning of each milking event, 30–45 s were required for the CH 4 and CO 2 concentrations to reach stable high levels, and most events lasted at least 5 min. Accordingly, a restricted recording window of 30–300 s, which excluded the initial unstable period while retaining data from the majority of milking events, was established. Gas concentrations significantly exceeding the baseline were selected as responses to ensure that only data obtained when the cows’ heads were sufficiently close to the sniffer air inlets were included. The mean value of the selected records was used as the response phenotype for each milking event. The concentration phenotypes showed moderate to high repeatability, but the CH 4 :CO 2 ratio had only moderate repeatability. The pipeline developed in this study enables the effective extraction of baseline-adjusted emissions phenotypes from sniffer data obtained in milking robots.